MATHEMATICAL BIOSCIENCES doi:10.3934/mbe.

2017004
AND ENGINEERING
Volume 14, Number 1, February 2017 pp. 45–66

ON THE MATHEMATICAL MODELLING OF

TUMOR-INDUCED ANGIOGENESIS

Luis L. Bonilla
G. Millán Institute, Fluid Dynamics, Nanoscience and Industrial Mathematics
Universidad Carlos III de Madrid
28911 Leganés, Spain

Vincenzo Capasso∗
ADAMSS, Universitá degli Studi di Milano
20133 MILANO, Italy

Mariano Alvaro, Manuel Carretero and Filippo Terragni

G. Millán Institute, Fluid Dynamics, Nanoscience and Industrial Mathematics
Universidad Carlos III de Madrid
28911 Leganés, Spain

Abstract. An angiogenic system is taken as an example of extremely com-

plex ones in the field of Life Sciences, from both analytical and computational
points of view, due to the strong coupling between the kinetic parameters
of the relevant branching - growth - anastomosis stochastic processes of the
capillary network, at the microscale, and the family of interacting underlying
biochemical fields, at the macroscale. To reduce this complexity, for a concep-
tual stochastic model we have explored how to take advantage of the system
intrinsic multiscale structure: one might describe the stochastic dynamics of
the cells at the vessel tip at their natural microscale, whereas the dynamics
of the underlying fields is given by a deterministic mean field approximation
obtained by an averaging at a suitable mesoscale. But the outcomes of relevant
numerical simulations show that the proposed model, in presence of anasto-
mosis, is not self-averaging, so that the “propagation of chaos” assumption
cannot be applied to obtain a deterministic mean field approximation. On the
other hand we have shown that ensemble averages over many realizations of the
stochastic system may better correspond to a deterministic reaction-diffusion
system.

1. Introduction. In biology and medicine we may observe a wide spectrum of

self-organization phenomena. This may happen at any scale; from the cellular scale
of embryonic tissue formation, wound healing or tumor growth, and angiogenesis,
to the much larger scale of animal grouping. Such phenomena are usually explained
in terms of a collective behavior driven by “forces”, either external and/or internal,
acting upon individuals (cells or organisms). In most of these organization phe-
nomena, randomness plays a major role; here we wish to address the issue of the

2010 Mathematics Subject Classification. Primary: 60G57, 60H10, 60H30, 60B10, 82C31,
92B05; Secondary: 60K35, 65C20, 65C35.
Key words and phrases. Angiogenesis, stochastic differential equations, birth and death pro-
cesses, growth processes, mean field approximation, hybrid models, propagation of chaos, ensemble
average.
∗ Corresponding author: Vincenzo Capasso.

45
46 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

relevance of randomness as a key feature for producing nontrivial organization of

biological structures (see [8] for a general discussion on this topic). As a working
example we offer a review of papers by the same authors in which tumor-driven an-
giogenesis has been analyzed [10], [4], [41]. In this case cells organize themselves as a
capillary network of vessels, the organization being driven by a family of underlying
fields, such as nutrients, growth factors and alike [22, 25, 13, 19, 21, 15, 14].
Actually an angiogenic system is extremely complex due to its intrinsic multiscale
structure. When modelling such systems, we need to consider the strong coupling
between the kinetic parameters of the relevant microscale branching and growth
stochastic processes of the capillary network and the family of interacting macroscale
underlying fields. Capturing the keys of the whole process is still an open problem
while there are many models in the literature that address some partial features of
the angiogenic process [2, 36, 30, 31, 27, 37, 38, 35, 10, 39, 34, 20].
The importance of using an intrinsically stochastic model at the microscale to
describe the generation of a realistic vessel network has been the subject of a series
of papers by one of the present authors [8, 11].
Typical models treat vessel cells on the extracellular matrix as discrete objects,
and different cell processes like migration, proliferation, etc. occur with certain
probabilities. The latter depend on the concentrations of certain chemical factors
which satisfy systems of reaction-diffusion equations (RDEs) [2, 24, 42, 34].
Viceversa, the RDEs for such underlying fields contain terms that depend on
the spatial distribution of vascular cells. As a consequence, a full mathematical
model of angiogenesis consists of the (stochastic) evolution of vessel cells, coupled
with a system of RDEs containing terms that depend on the distribution of ves-
sels. The latter is random and therefore the equations for the underlying fields are
random RDEs, which are supposed to drive the kinetic parameters of the stochas-
tic geometric processes of birth (branching), growth (vessel extension), and death
(anastomosis).
This strong coupling leads to an highly complex mathematical problem from both
analytical and computational points of view. A possibility to reduce complexity is
offered by the so called hybrid models, which exploit the natural multiscale nature
of the system.
The idea consists of approximating the random RDEs by deterministic ones, in
which the microscale (random) terms depending on cell distributions are replaced by
their (deterministic) mesoscale averages. In this way only a simple stochasticity of
the geometric processes of branching - vessel extension - anastomosis is kept, driven
now by deterministic kinetic parameters depending upon the above mentioned mean
field approximation of the concentrations of the relevant fields [10, 28].
Our analysis has shown that nontrivial problems arise when deriving determinis-
tic equations for the mean spatial densities of the relevant stochastic entities mod-
elling the vessels’ evolution at the microscale.
In the literature there are examples of rigorous derivations of mean field equations
of stochastic particle dynamics [29, 40, 17, 6]. However, to the best of the authors’
knowledge, the kind of models considered here have not yet been studied and require
further investigation.
In angiogenesis, the leading mechanisms driven by the underlying fields consist of
vessel branching, elongation and anastomosis. This last one has been the major and
critical addition to the model proposed in [10]. In [4] anastomosis has been modelled
as a death process of a tip that encounters an existing vessel and is therefore coupled
 TUMOR-INDUCED ANGIOGENESIS 47

with the density of the full vessel network. In order to cope with anastomosis, the
vessel network has been modelled as a stochastic geometric process of Hausdorff
dimension one, as opposed to the system of tips which form a usual stochastic
particle system of Hausdorff dimension zero.
In [4] we have been able to derive (at least formally) a mean field equation for the
spatial density of tips, as a function of spatial location and velocity. This equation
is a parabolic integrodifferential equation of a Fokker-Planck type, having a source
term and a noninvertible diffusion matrix; it is second order in the derivatives with
respect to the velocities, and first order in the derivatives with respect to the position
coordinates. Apparently, together with the mean field equations for the underlying
fields, we have thus found an independent (deterministic) integrodifferential system
whose solution can provide the required (deterministic) kinetic parameters, which
drive the stochastic system for the tips, eventually leading to the stochastic vessel
network, at the microscale.
Mean field equations that follow from the “propagation of chaos” assumption
(law of large numbers) (see [29], [40], and references therein) are quite convenient
as they hold for any given realization of the underlying self-averaging stochastic
processes, but this requires a large number of tips at any location in the phase
space, and at any time.
Unfortunately anastomosis is responsible for a significant decay of the number
of tips, so to make inapplicable laws of large numbers on a single realization of
the stochastic process (a replica of the system), according to the “propagation of
chaos”.
However by re-examining the derivation given in [4], in [41] we have noticed
that the same deterministic description holds for vessel tip densities calculated by
averaging over replicas. This is close to J.W. Gibbs’ original ensemble average
interpretation of equilibrium statistical mechanics [23], except that, of course, our
system is always very far from equilibrium. In either cases, though with a different
interpretation of the solution, the deterministic description consists of a reaction-
diffusion equation for the tumor angiogenic factor (TAF) concentration coupled to
a Fokker-Planck type equation for the vessel tip density. To conclude, randomness
cannot be avoided; the deterministic description represents the evolution of an ideal
“mean” angiogenesis system and the evolution of a particular replica may deviate
significantly from it. These deviations are important and deserve further study, but
this is outside the scope of the present paper (see [8]).
The paper is organized as follows. Section 2 describes how our stochastic model
treats vessel branching, extension and anastomosis. In the subsequent three sections
we present the mathematical ingredients to be used in the sequel. In Section 6 we
derive the mean field approximation based on the propagation of chaos assumption.
In Section 7 a discussion is presented based on the outcomes of the numerical
simulations of the stochastic model as opposed to the mean field model. In Section 8
the ensemble average approach is presented to derive an equation of Fokker-Planck
type for the density of vessel tips and the TAF’s RDE. In Section 9 numerical
simulations of the deterministic averaged model are presented as opposed to the
stochastic model. Finally Section 10 contains our conclusions.

2. The mathematical model. Based on the above discussion, the main features
of the process of formation of a tumor-driven vessel network that we have considered
are (see [18, 30, 10])
48 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

i) vessel branching;
ii) vessel extension;
iii) chemotaxis in response to a generic tumor angiogenic factor (TAF), released
by tumor cells;
iv) haptotactic migration in response to fibronectin gradient, emerging from the
extracellular matrix and through degradation and production by endothelial
cells themselves;
v) anastomosis, the coalescence of a capillary tip with an existing vessel.

Let N0 denote the initial number of tips, N (t) the number of tips at time t. Each
particle tip is characterized by its space Xi (t) and velocity vi (t) coordinates, so that
the whole process is characterized by the stochastic processes {(Xi (t), vi (t)), i =
1, ..., N (t), t ∈ R+ }. We model the capillary network of endothelial cells X(t) as the
union of all random trajectories representing the extension of individual capillary
tips from the (random) time of birth (branching) T i , to the (random) time of death
(anastomosis) Θi ,
N (t)
[
X(t) = {Xi (s), T i ≤ s ≤ min{t, Θi }}, (1)
i=1

giving rise to a stochastic network. It is clear that Xi (t) is a random closed set of
Hausdorff dimension zero, while X(t) is a random closed set of Hausdorff dimension
one.
We may describe both random sets by means of random distributions á la Dirac-
Schwarz [12]; δXi (t) (x) is the random distribution of the tip Xi (t) localized at point
x, for i = 1, . . . , N (t); δX(t) (x) is the random distribution of the whole network
X(t) localized at point x.
For convenience of the reader, we remind here that, under sufficient regularity
assumptions on a random set Ξ, it may admit a mean density λΞ , given by the
expectation of the random distribution [12] [1]
λΞ (x) = E[δΞ ](x). (2)

2.1. The underlying fields. TAF, fibronectin and matrix degrading enzymes ac-
tivate the migration of endothelial cells.
TAF diffuses, and it decreases where endothelial cells are present; we will adopt
the following model for the TAF evolution, according to which “consumption” (ac-
tually molecular binding) is due to the additional endothelial cells producing ves-
sels’ extension. Consumption is then taken proportional to the velocity vi , i =
1, . . . , N (t), of the tips, hence
∂
C(t, x) = d0 δA (x) + d2 4C(t, x)
∂t
N (t)
1 X i
− ηC(t, x) |v (t)|(KN ∗ δXi (t) )(x). (3)
N i=1
Parameters d2 , η > 0 denote the diffusivity and the rate of consumption, respec-
tively, while d0 represents the rate of production by a source located in a region
A ⊂ Rd , modelling e.g. a tumor mass. Later we will include the production term in
the evolution equation for C(t, x) only via the boundary conditions, meaning that
the tumor is located at the boundary of the relevant domain.
 TUMOR-INDUCED ANGIOGENESIS 49

Model (3) considers a dependence upon the (mollified) empirical distribution of

the variation in length of the existing vessels, per unit time.
The parameter N represents a scale parameter, corresponding to the order of
magnitude of the number of vessels in the network, so that the action of each
existing vessel is reduced accordingly. Convolution with the kernel KN (x) provides
a mollified version of the relevant random distributions; from a modelling point
of view this may correspond to a nonlocal reaction with the relevant modelling
fields (see e.g. [38]). Correspondingly, the mollifier kernel KN is chosen such that
lim KN (x) = δ0 (x).
N →∞
Specific choices about the dependence of the kernel KN upon N may allow the
convergence, for N tending to infinity, of the mollified random distributions to the
corresponding mean densities, thanks to suitable laws of large numbers.
Fibronectin is bound to the extracellular matrix and does not diffuse [3]. Degra-
dation of fibronectin, characterized by a coefficient ζ1 , depends on the concentration
of matrix degrading enzyme(MDE), produced by the cells [35].
N (t)
∂ 1 X
f (t, x) = −ζ1 m(t, x)f (t, x) + ζ2 (KN ∗ δXi (t) )(x). (4)
∂t N i=1
The MDE, once produced at a rate ν1 , diffuses locally with a diffusion coefficient
1 , and is spontaneously degraded at a rate ν2 .
N (t)
∂ 1 X
m(t, x) = 1 4m(t, x) − ν2 m(t, x) + ν1 (KN ∗ δXi (t) )(x). (5)
∂t N i=1

All these (random) partial differential equations are subject to suitable boundary
and initial conditions.
For the sake of technical simplification, in the sequel we will not consider the last
two underlying fields f and m.

3. The relevant empirical measures. Let us assume that, at any time t ≥ 0,

the number of tips, N (t), and the number of vessels per unit volume is of the
same order O(N ), where N is a scale parameter; at first we shall consider the
“propagation of chaos” approach according to which N can be taken as a sufficiently
large positive integer, uniformly in space and time; the main issue of this paper is
that this assumption may be false, as shown by the numerical simulations of the
model presented here (see Section 7, and the following ones).
There are two fundamental random measures describing the system at any time
t ≥ 0. The global random empirical measure QN of the joint process {(Xk (t), vk (t)),
k = 1, . . . , N (t), t ∈ R+ } is defined as
N (t)
1 X
QN (t) := (Xk (t),vk (t)) . (6)
N
k=1

Here (x,v) is the usual Dirac measure on BRd ×Rd .

Consequently, the random empirical measure of the process {(Xk (t)), k = 1, . . . ,
N (t), t ∈ R+ } will be given by
N (t)
1 X
TN (t) = Xk (t) = QN (t)(· × Rd ). (7)
N
k=1
50 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

4. The dynamics.

4.1. Tip branching. Two kinds of branching have been identified; either from a
tip or from a mature vessel; here for the sake of simplicity, we shall consider only
tip branching.
The birth process of new tips can be described in terms of a marked point process
(see e.g. [5]), by means of the random measure Φ on BR+ ×Rd ×Rd such that, for any
t ≥ 0 and any B ∈ BRd ×Rd ,
Z tZ
Φ((0, t] × B) := Φ(ds × dx × dv), (8)
0 B
where Φ(ds × dx × dv) is the random variable that counts those tips born from an
existing tip during times in (s, s + ds], with positions in (x, x + dx], and velocities
in (v, v + dv].
We assume that the probability that a tip branches from one of the N (t) ex-
isting ones during an infinitesimal time interval (t, t + dt] is function of the TAF
concentration C(t, x)
N (t)
X
prob(N (t + dt) − N (t) = 1|Ft− ) = α(C(t, Xi (t))) dt, (9)
i=1

where α(C) is a non-negative function; for example, we may take

C
α(C) = α1 , (10)
CR + C
where CR is a reference density parameter [10].
Here Ft− denotes the history of the whole process up to time t− .
As a technical simplification, we will further assume that whenever a tip located
in x branches, the initial value of the state of the new tip is (XN (t)+1 , vN (t)+1 ) =
(x, v0 ), where v0 is a non random velocity. Later, as in [41], a newly created tip
will be taken to assume an initial velocity selected out of a normal distribution with
mean v0 , and some variance.
We then claim that the compensator of the random measure Φ(ds × dx × dv) is
α(C(s, x))δv0 (v) N QN (s)(d(x, v)) ds. (11)

4.2. Vessel extension. We describe vessel extension by the Langevin system

dXk (t) = vk (t) dt
dvk (t) = −k vk (t) + F C(t, Xk (t)) dt + σ dWk (t),


(12)
k k
for all those k = 1, . . . , N (t) for which T < t < Θ . Besides the friction force, there
is a force due to the underlying TAF field C(t, x) [30, 35]:
d1
F(C) = ∇x C. (13)
(1 + γ1 C)q
4.3. Anastomosis. When a vessel tip meets an existing vessel it joins it at that
point and time and it stops moving. This process is called tip-vessel anastomosis.
As in the case of the branching process, we have modelled this process via a
marked counting process; anastomosis is modelled as a “death” process.
Let Ψ denote the random measure on BR+ ×Rd ×Rd such that, for any t ≥ 0, and
any B ∈ BRd ×Rd ,
 TUMOR-INDUCED ANGIOGENESIS 51

Z tZ
Ψ((0, t] × B) := Ψ(ds × dx × dv) (14)
0 B
where Ψ(ds × dx × dv) is the random variable counting those tips that are absorbed
by the existing vessel network during time (s, s + ds], with position in (x, x + dx],
and velocity in (v, v + dv].
Thanks to the choice of a Langevin model for the vessels extension, we may
assume that the trajectories are sufficiently regular and have an integer Hausdorff
dimension 1.
Hence [12] the random measure
A ∈ BRd 7→ H1 (X(t) ∩ A) ∈ R+ (15)
admits a random generalized density δX(t) (x) with respect to the usual Lebesgue
measure on Rd such that, for any A ∈ BRd ,
Z
H1 (X(t) ∩ A) = δX(t) (x)dx. (16)
A

Given the definition of X(t) as the union of all trajectories up to time t, the
Hausdorff measure associated with the stochastic network X(t) can be expressed in
terms of the occupation time of a region A ∈ BRd by tips existing up to time t > 0
(see [32], [26])
Z t N (s)
X
H1 (X(t) ∩ A) = γ ds IA (Xi (s))
0 i=1
Z t N (s)
X
= γ ds Xi (s) (A), (17)
0 i=1

where IA (x) = 1 if x ∈ A, 0 otherwise, and x denotes the usual Dirac measure

associated with the point x. γ is a suitable dimensional constant taking into account
an average velocity of growth of the vessels; it can be used to fit the model with
real data.
The latter has the Dirac delta as its generalized derivative with respect to the
usual Lebesgue measure on Rd . Hence,
Z t N (s)
X
δX(t) (x) = γ ds δXi (s) (x). (18)
0 i=1

We assume that the probability per unit time that a typical tip Xk (t) meets the
existing vessel network X(t) (and “dies”) is a linear function of the scaled random
distribution N −1 δX(t) of the stochastic network at point Xk (t) and time t, so that
the probability that an active tip dies during an infinitesimal time interval is given
by
N (t)
X
prob(N (t + dt) − N (t) = −1|Ft− ) = N −1 (KN ∗ δX(t) )(Xi (t))dt. (19)
i=1

We then claim that the compensator of the random measure Ψ(ds × dx × dv) is
N −1 (KN ∗ δX(s) )(x)N QN (s)(d(x, v))ds. (20)
52 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

5. Complexity. Our stochastic model is thus described by a set of Ito stochastic

differential equations (SDEs), a marked point process describing tip branching (a
birth process), and a marked point process describing anastomosis (a death process).
The latter process depends on the past history of a given realization of the overall
stochastic process. In addition, the TAF concentration is itself a random process
since it depends on the stochastic evolution of the tips as indicated by Equation
(3). Hence C(t, x) is a random spatial field, which is supposed to drive the kinetic
parameters of branching and extension of vessels (see Equations (9), and (12)); as a
consequence the fully stochastic system described in the previous sections results to
be of high complexity, both analytical and computational, especially in presence of
a large vessel network. In most of the current literature a reduction of the complex-
ity is reached by taking a deterministic mean field approximation of the stochastic
reaction term in the evolution equation of the underlying fields. In this way the
stochastic processes of branching and extension are driven by deterministic param-
eters. The usual assumption is the “propagation of chaos”; as we had anticipated
in Section 3 this is possible only if one can claim that the scale parameter N can
be taken as a sufficiently large positive integer, uniformly in space and time. Next
section is devoted to that approach.

6. Mean field approximation. By Itô’s formula (see e.g. [9], p. 270), we may
obtain the evolution equation of the random measure (QN (t))t∈R+ as follows. For
any test function g ∈ Cb (Rd × Rd ),
Z Z
g(x, v) QN (t)(d(x, v)) = g(x, v) QN (0)(d(x, v))
Z tZ
+ v · ∇x g(x, v)QN (s)(d(x, v)) ds
0
Z tZ
+ [F(C(s, x)) − kv]·∇v g(x, v)QN (s)(d(x, v)) ds
0
Z tZ 2
σ
+ ∆v g(x, v)QN (s)(d(x, v))ds
0 2
Z tZ
+ α(C(s, x))δv0 (v)g(x, v)QN (s)(d(x, v)) ds
0
Z tZ
1
− (KN ∗ δX(s) )(x)g(x, v)QN (s)(d(x, v)) ds
0 N
+M̃N (t), (21)
where
M̃N (t) = M̃1,N (t) + M̃2,N (t) + M̃3,N (t),
with
t N (s)
1 X
Z
M̃1,N (t) = ∇v g(Xk (s), vk (s)) · dWk (s), (22)
0 N
k=1

Z tZ
M̃2,N (t) = g(x, v)[ΦN (ds × dx × dv)
0
−α(C(s, x))δv0 (v)QN (s)(d(x, v))ds], (23)
 TUMOR-INDUCED ANGIOGENESIS 53

Z tZ
M̃3,N (t) = g(x, v)[ΨN (ds × dx × dv)
0

1
− (KN ∗ δX(s) )(x)QN (s)(d(x, v))ds . (24)
N
All (M̃1,N (t))t∈R+ , (M̃2,N (t))t∈R+ , and (M̃3,N (t))t∈R+ are zero mean martingales,
so that (M̃N (t))t∈R+ is itself a zero mean martingale.
By Doob’s inequality
P
M̃N (t) −→ 0.
N →∞
This is the substantial reason of the possible deterministic limiting behavior of
the process, as N → ∞.
As a consequence, if we suppose that indeed, for a large value of the scale pa-
rameter N, we may claim
QN (t)(d(x, v)) → Q∞ (t)(d(x, v)) = p(t, x, v)dx dv. (25)
1
Assuming that (25) holds true, the scaled stochastic distribution of vessels (KN ∗
N
δX(t) )(x) can be itself approximated by the mean value
Z t
λ(t, x) = E[δX(t) ](x) = γ p̃(s, x) ds, (26)
0
where Z
p̃(t, x) = E[δXi (t) ](x) = p(t, x, v0 ) dv0 (27)

is the marginal density of p(t, x, v).

A formal justification of the above is the following one. According to a previous
discussion
Z t N (s)
1 1 X
δX(t) (x) = γ ds δ i (x).
N 0 N i=1 X (s)
In the mean field approximation we are assuming that a “law of large numbers” may
N (s)
1 X
be applied, so that δ i (x) is an approximation of p̃(s, x), and consequently,
N i=1 X (s)
Z t
1
(KN ∗ δX(t) )(x) is an approximation of γ p̃(s, x)ds.
N 0
For objects of Hausdorff dimension 0, as they are (Xi (t), vi (t)), it is well known
that the only request for stating that their expected value p(t, x, v) is a classical
function, is that they are absolutely continuous random vectors, and this is the case
for solutions of SDEs. On the other hand, the possibility that the generalized func-
tion δX(t) describing the vessel network, of Hausdorff dimension 1, is an absolutely
continuous random object is based on the crucial assumption of a Langevin model
for the vessels’ extension (see System (12)). Indeed under this choice we may state
that almost all trajectories are rectifiable; a pure Brownian motion model would
have led to trajectories that are not rectifiable though continuous (the mathematical
theory of random densities is presented in [12]).
Finally the consumption term in (3) is an approximation of the flux density
Z
j(t, x) = v0 p(t, x, v0 ) dv0 , (28)
54 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

so that the evolution equation (3) may be approximated by its (deterministic) mean
field version
∂ e
C(t, x) = d2 ∆x C(t,
e x) − η C(t,
e x)|j(t, x)|. (29)
∂t
TAF injection from the tumor is realized as a nonzero flux boundary condition for
this equation [4], instead of including it explicitly as in (3).
Based on the above discussion, one might claim that the evolution equation for
the density p(t, x, v) is the following one

∂ α1 C(t,
e x)
p(t, x, v) = p(t, x, v)δv0 (v)
∂t CR + C(t,
e x)
Z t
−γp(t, x, v) p̃(s, x) ds
0
−v · ∇x p(t, x, v)
+k∇v · (vp(t, x, v))
" #
∇x C(t,
e x)
−d1 ∇v · p(t, x, v)
e x)]q
[1 + γ1 C(t,
σ2
+ ∆v p(t, x, v). (30)
2
The coupled system (29), (30) is subject to suitable boundary and initial condi-
tions, to be discussed later.
This approach is called “hybrid”, since we have substituted all stochastic un-
derlying fields by their “mean” counterparts; most of the current literature could
now be reinterpreted along these lines. Indeed, one should check that the hybrid
system is fully compatible with a rigorous derivation of the evolution for the ves-
sel densities. Nonlinearities in the full model are a big difficulty in this direction;
a rigorous derivation of the convergence results requires a nontrivial mathematical
analysis, which is under investigation; for this it is instrumental a proof of existence,
uniqueness and sufficient regularity of the solution of the mean field equations (see
[29], and [6]). We wish to stress that anyhow substituting mean geometric densities
of tips or of full vessels to the corresponding stochastic quantities leads to an ac-
ceptable coefficient of variation (percentage error) only when a law of large numbers
can be applied, i.e. whenever the relevant numbers per unit volume are sufficiently
large; otherwise stochasticity cannot be avoided, and in addition to mean values,
the mathematical analysis and/or simulations should provide confidence bands for
all quantities of interest (see e. g. [7]). An interesting case in this direction is
discussed in [11].

7. Numerical simulations. Figure 1 shows the outcome of a numerical simulation

of the fully stochastic model. The values of all parameters are discussed in [4] and
[41]; in particular all kernels have been taken as Gaussians centered at 0 . Tips
proliferate by branching but they tend to crowd in a relatively narrow region due to
chemotaxis. The number of active tips is kept below a hundred by anastomosis and,
as a result, there never are enough tips for the law of large numbers to produce self-
averaging of realizations. In fact, the fluctuations of tip density or velocity observed
in simulations are not small at any time. Thus a deterministic description of vessel
 TUMOR-INDUCED ANGIOGENESIS 55

tip density based on self-averaging due to the law of large numbers does not seem
correct.

Time = 8 h – Number of tips = 28 Time = 16 h – Number of tips = 47 Time = 24 h – Number of tips = 53

0.5 0.5 0.5

0.4 0.4 0.4

0.3 0.3 0.3

0.2 0.2 0.2

0.1 0.1 0.1

y/L 0 y/L 0 y/L 0

−0.1 −0.1 −0.1

−0.2 −0.2 −0.2

−0.3 −0.3 −0.3

−0.4 −0.4 −0.4

−0.5 −0.5 −0.5

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

x/L x/L x/L

1.2 1.2 1.4

1 1 1.2
1
0.8 0.8
0.8
0.6 0.6
0.6
0.4 0.4
0.4
0.2 0.2
0.2
0 0 0
−0.2 −0.2 −0.2
1 1 1
0.5 1 0.5 1 0.5 1
0.8 0.8 0.8
0 0 0
y/L −0.5 0.4
0.6
y/L −0.5 0.4
0.6
y/L −0.5 0.4
0.6

−1 0
0.2
x/L −1 0
0.2
x/L −1 0
0.2
x/L

Figure 1. Upper panels: Snapshots of the vessel network inside a

central square of side L at three different times. The level curves of
the TAF density C(t, x) are also depicted. Middle panels: Density
plots of the TAF density at the same times as in the upper panels
showing its consumption as the vessel tips advance. Lower panels:
Surface plots of the x-component of the chemotactic force at the
same times showing how it pushes the tips toward the tumor.

However, there is another interpretation of the numerical simulations for which

a deterministic description is correct. While the vessel network may look different
for different replicas of the stochastic process, tip densities associated to averages
over replicas are described by Equation (30) [41]. How is this possible? Let N be
the number of independent replicas (realizations) of the angiogenic process having
the same initial number of vessel tips but otherwise random initial conditions. For
any replica ω of the stochastic simulation, let us define the stochastic distribution
of tips per unit volume in the (x, v) phase space by
N (t,ω)
X
∗
QN (t, x, v, ω) = δXi (t,ω) (x) δvi (t,ω) (v). (31)
i=1

Here the number of tips at time t, N (t, ω), may be different for different replicas.
Similarly at time t, the stochastic distribution of tips per unit volume in physical
space is given by
N (t,ω)
X
Q̃∗N (t, x, ω) = δXi (t,ω) (x), (32)
i=1
56 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

and
Z t NX
(s,ω)
δX(t,ω) (x) := γ δXi (s,ω) (x)ds (33)
0 i=1
represents the concentration of all vessels per unit volume in physical space, at time
t, i.e., the vessel network.
By following a similar approach as in our previous paper [4], we may then ob-
tain the weak formulation of the stochastic evolution of Q∗N (t, x, v, ω), by following
similar steps as in Equation (21):
Z Z
∗
g(x, v) QN (t, x, v)dxdv = g(x, v) Q∗N (0, x, v)dxdv
Z tZ
+ v · ∇x g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ
+ [F(C(s, x)) − kv]·∇v g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ 2
σ
+ ∆v g(x, v)Q∗N (s, x, v)dxdvds
0 2
Z tZ
+ α(C(s, x))δv0 (v)g(x, v)Q∗N (s, x, v)dxdv ds
0
Z tZ
− δX(s) (x)g(x, v)Q∗N (s, x, v)dxdv ds
0
+ M N (t), (34)
where we have omitted the variable ω not to encumber the formulas. Here M N (t) is
again a zero mean martingale that collects all source of randomness of the system,
g(x, v) is a smooth test function.
By mimicking the typical kernel density estimation approach in Statistics (see
e.g. [33], page 489), we introduce the (random) empirical distribution of tips per
unit volume in the (x, v) phase space,
N
1 X
pN(t, x, v) = (KN ∗ Q∗N (t, ·, ·, ω))(x, v). (35)
N ω=1
Here the mollifying kernel KN (x, v) tends to a Dirac delta function δ0,0 as N → ∞.
Correspondingly, the empirical distribution of tips per unit volume in the physical
space and the vessel tip flux are,
N
1 X e ∗N (t, ·, ω))(x),
p̃N (t, x) = (KN ∗ Q (36)
N ω=1
N N (t,ω)
1 X X i
jN (t, x) = v (t, ω)(KN ∗ δXi (t,ω) )(x), (37)
N ω=1 i=1
respectively. We now conjecture that the following limit exists
p(t, x, v) = lim pN (t, x, v), (38)
N →∞

and that p(t, x, v) is the deterministic distribution of tips per unit phase space
volume. The proof of this statement is not trivial (and out of the scope of this
paper), as the usual assumptions on the kernel density estimation (see page 489 of
 TUMOR-INDUCED ANGIOGENESIS 57

[33]) may not apply. Correspondingly the deterministic distribution of tips per unit
volume in the physical space should exist as the following limit

p̃(t, x) = lim p̃N (t, x). (39)

N →∞

Finally, we may obtain the deterministic version of the vessel tip flux as

j(t, x) = lim jN (t, x). (40)

N →∞

Figure 2. Density plots of the marginal tip density p̃(t, x, y) cal-

culated from (36) with N = 50 replicas for the initial time and the
same times as in Figure 1. The panels show how tips are created
at x = 0 and march toward the tumor at x = L.

Figure 2 shows the marginal tip density p̃(t, x, y) ≈ p̃N (t, x, y) calculated from
(36) with N = 50 replicas at the times represented in Figure 1. The tips proliferate
after a few hours and reach a high value by branching onto the free space ahead
of them. Influenced by chemotaxis, the marginal tip density thickens about the x
axis and it forms a lump that advances toward the tumor. Behind the lump, the
density drops to a low value as few active tips remain. As shown by Figure 1, the
network of vessels is quite dense in the wake of the tips and therefore anastomosis
reduces enormously the number of active tips there. They are numerous only at
the leading part of the lump where free space is available. Thus the definition
of marginal tip density based on ensemble average over replicas provides a better
deterministic description of angiogenesis than the density based on the law of large
numbers. There are no tips or very few ones in large regions of the physical space.
58 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

8. Ensemble average description. On the basis of the above convergence as-

sumptions and considering
N
1 X
lim M N (t, ω) = 0,
N →∞ N
ω=1

we may expect that the ensemble average of the stochastic equation (34) tends in
its strong form to the same equation of Fokker-Planck type as (30):
∂ α1 C(t, x)
p(t, x, v) = p(t, x, v)δv0 (v)
∂t CR + C(t, x)
Z t
− γp(t, x, v) p̃(s, x) ds − v · ∇x p(t, x, v)
0
 
∇x C(t, x)
+ k∇v · [vp(t, x, v)] − d1 ∇v · p(t, x, v)
[1 + γ1 C(t, x)]q
σ2
+ ∆v p(t, x, v). (41)
2
Here the marginal vessel tip density,
Z
p̃(t, x) = p(t, x, v0 ) dv0 , (42)

is the marginal density of p(t, x, v). As an additional consequence of the above,

Equation (41) for p(t, x, v) is coupled with a deterministic reaction-diffusion equa-
tion for the TAF concentration,
∂
C(t, x) = d2 ∆x C(t, x) − η C(t, x)|j(t, x)|, (43)
∂t
where j(t, x) is the ensemble-averaged current density (flux) vector at any point x
and any time t ≥ 0,
Z
j(t, x) = v0 p(t, x, v0 ) dv0 . (44)

A crucial point in the derivation of (41) as the ensemble average of Equation (34)
is the approximation of nonlinear terms in the latter, e.g. the anastomosis term.
Had the law of large numbers been applicable (on the single replica, according to
the propagation of chaos assumption), the mean values of nonlinear terms could be
factorized. With the ensemble average description, factorization of nonlinear terms
requires further investigation, and this is out of the scope of the present paper. By
assuming factorization occurs, we may expect that the ensemble average of Equation
(34) tends in its strong form to (41).
By an abuse of notations, we are using in Equation (43) the same letter C for
the deterministic counterpart of the TAF field, too.
We need to stress here that the p(t, x, v) appearing in Equation (30) represents
a scaled density in phase space, the integral of which over the whole phase space
is N (t)/N , where N (t) is the mean number of tips at time t. This number coin-
cides with the number of tips in a single realization of our supposed self-averaging
stochastic process. In contrast to this, the p(t, x, v) appearing in Equation (41)
represents the true concentration per unit volume of phase space.
For appropriate initial and boundary data, it is possible to prove that (41) and
(43) have a unique smooth solution [16].
 TUMOR-INDUCED ANGIOGENESIS 59

8.1. Boundary and initial conditions. We have solved the system of equations
(41) and (43) in a two dimensional strip geometry using the initial and boundary
conditions introduced in [4]. The strip is Ω = [0, L] × R ⊂ R2 , its left boundary
Ω0 = (0, y), y ∈ R, is the primary vessel issuing new vessels, and ΩL = (L, y),
y ∈ R, represents the tumor which is a source of the TAF C. Let c1 (y) be the TAF
flux emitted by the tumor at x = L. As boundary conditions for the TAF we have
taken
∂ ∂ c1 (y)
C(t, 0, y) = 0, C(t, L, y) = , (45)
∂x ∂x d2
and C → 0 as |y| → ∞. We have used a Gaussian as the initial condition for the
TAF
2
/c2 +y 2 /b2 ]
C(0, x, y) = 1.1 CR e−[(x−L) , (46)
for appropriate b and c.
As boundary conditions for the tip density we have taken
k|v−v0 |2
e− σ2
p+ (t, 0, y, v, w) = R R k|v0 −v0 |2
∞ ∞
0 −∞
v 0 e− σ2dv 0 dw0
 Z 0 Z ∞ 
0 − 0 0 0 0
× j0 (t, y)− v p (t, 0, y, v , w )dv dw , (47)
−∞ −∞
k|v−v0 |2

− e− σ2
p (t, L, y, v, w) = R R k|v0 −v0 |2
0 ∞
−∞ −∞
e− σ2dv 0 dw0
 Z ∞Z ∞ 
+ 0 0 0 0
× p̃(t, L, y)− p (t, L, y, v , w )dv dw , (48)
0 −∞
p(t, x, v) → 0 as |v| → ∞, (49)
+ −
where p = p for v > 0 and p = p for v < 0, v = (v, w). These phenomenological
conditions give the tip density for velocities entering the slab region in terms of the
tip density for velocities exiting the slab region. They are constructed so that they
give the marginal tip density at x = L and the tip flux density at x = 0, which is
[4]
v0 L
j0 (t, y) = p 2 α(C(t, 0, y)) p(t, 0, y, v0 , w0 ), (50)
v0 + w02
for the vector velocity v0 = (v0 , w0 ). Improving the boundary conditions requires a
separate and detailed work comparing stochastic and deterministic descriptions of
the angiogenesis process.
Finally as initial condition for the tip density we have taken
2 2 N
2e−x /lx −|v−v0 |2 /σv2 X 0
1 −|y−yi |2 /ly2
p(0, x, y, v, w) = e √ e . (51)
π 3/2 lx σv2 i=1
πly

As lx and ly tend to zero, (51) becomes

N0
X
p(0, x, y, v) = 2δvσ0v (v)δ0 (x) δyi (y), (52)
i=1

where δvσ0v (v) is the mollified version of the usual Dirac delta by a Gaussian kernel.
This initial condition corresponds to the following initial condition for the stochastic
60 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

process: There are N0 equally spaced initial tips at x = 0, with vertical positions
yi equally spaced on the interval [−Ly , Ly ], whose initial velocities are normally
distributed about v0 with standard deviation σv .

9. Numerical results and comparison with the stochastic simulations.

We have used the parameter values that have been extracted from experiments as
explained in [4]. The anastomosis coefficient γ had been given an arbitrary value in
[4], whereas we estimate it here so as to get a good agreement between simulations
of the stochastic equations and solutions of the deterministic equations (see [41]).
We have suitably nondimensionalized the governing equations of our model, (41)
and (43), according to the units given in [4], thereby obtaining
Z t
∂p AC
= p δv0 (v) − Γp p̃(s, x) ds − v · ∇x p
∂t 1+C 0
  
δ ∇x C β
− ∇v · − βv p + ∆v p, (53)
(1 + Γ1 C)q 2
∂C
= κ ∆x C − χ C |j|. (54)
∂t
The dimensionless parameters appearing in these equations are defined in Table 1
(see [4], [41]). In the computations for the generalized function δv0 (v) that ap-
pears in (53) we have used the mollified version δv (v) with σv2 = σ 2 2 /k in (52) as
the variance in the Gaussian mollifying kernel, and obtained the nondimensional
function,
1 −|v|2 /2
δv (v) =
e . (55)
π2
We have used  = 0.08 in the numerical simulations.

δ β A Γ1 κ Γ χ
d1 CR kL α1 L d2 γ η
ṽ02 ṽ0 γ1 CR
ṽ03 ṽ0 Lṽ02L
1.5 5.88 22.42 0.145 1 0.0045 0.002
Table 1. Dimensionless parameters. ṽ0 = 40µm/h is a typical cell velocity.

The nondimensional boundary conditions for C are

∂C ∂C
(t, 0, y) = 0, (t, 1, y) = f (y), lim C = 0, (56)
∂x ∂x y→±∞

where f (y) = L c1 (Ly)/(CR d2 ) is a nondimensional flux, following from (45). We

2 2
have used c1 (y) = a e−y /b , with a = 5.5 × 10−27 mol/(m s), d2 = 10−13 m2 /s, and
b = 0.6 mm (b is about half the assumed tumor size). The initial condition for the
TAF (46) yields
2
L2 /c2 +y 2 L2 /b2 ]
C(0, x, y) = 1.1 e−[(x−1) , (57)
with b/L = 0.3, c/L = 1.5, whereas the nondimensional initial vessel density is
2 2 2 N
2Le−x L /lx −|v−v0 |2 /σv2 X 0
L −|y−yi |2 L2 /ly2
p(0, x, y, v, w) = 3/2 2
e √ e , (58)
π lx  i=1
πly
 TUMOR-INDUCED ANGIOGENESIS 61

with lx /L = 0.06 and ly /L = 0.08, that corresponds to N0 = 20 initial vessel tips.

In nondimensional form, the boundary conditions (47)-(48) for p become
2
e−|v−v0 |
p+ (t, 0, y, v, w) = R ∞ R ∞
0 −∞
v 0 e−|v0 −v0 |2 dv 0 dw0
 Z 0 Z ∞ 
× j0 (t, y)− v 0 p− (t, 0, y, v 0 , w0 )dv 0 dw0 (59)
−∞ −∞
for x = 0 and v > 0,
2
e−|v−v0 |
p− (t, 1, y, v, w) = R 0 R ∞
−∞ −∞
e−|v0 −v0 |2 dv 0 dw0
 Z ∞Z ∞ 
+ 0 0 0 0
× p̃(t, 1, y)− p (t, 1, y, v , w )dv dw (60)
0 −∞

for x = 1 and v < 0. Eq. (50) produces the nondimensional flux j0 :

C
j0 (t, y) = A v0 p(t, 0, y, v0 , w0 ) (61)
1+C
p
( v02 + w02 = |v0 |2 = 1 in nondimensional units).

Figure 3. Density plots of the marginal tip density calculated

from the deterministic description for the same times as in Figure
2.

In [4], we have shown the consistency of the deterministic model by depicting

TAF concentration, marginal tip density and overall network density at different
times. Figure 3 shows that the marginal tip density obtained by numerically solving
the deterministic equations (53)-(61) agrees quite well with the ensemble average
over 50 replicas of the stochastic process depicted in Figure 2.
62 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

Time = 0 h – Number of tips = 20 Time = 8 h – Number of tips = 43

1000 500

800 400

600 300

400 200

200 100

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L

Time = 16 h – Number of tips = 45 Time = 24 h – Number of tips = 50

400 500

400
300

300
200
200

100
100

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L

Time = 0 h – Number of tips = 20 Time = 8 h – Number of tips = 47

120 250

100 200
80
150
60
100
40

20 50

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L

Time = 16 h – Number of tips = 50 Time = 24 h – Number of tips = 56

250 200

200
150

150
100
100

50
50

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x/L x/L

Figure 4. Comparison of the marginal tip density at the x axis,

p̃(t, x, y = 0), as calculated from the deterministic description (up-
per panels) and from ensemble averages over 50 replicas of the
stochastic process (lower panels).

Figure 4 compares the deterministic and averaged stochastic descriptions of the

marginal tip density at the x axis, p̃(t, x, y = 0). The marginal density of the
vessel tips advances as a lump toward the tumor at x = L, and its projection
onto the x axis is a moving pulse. While both descriptions agree qualitatively,
there are quantitative discrepancies: the pulse maximum is larger for the solution
 TUMOR-INDUCED ANGIOGENESIS 63

of the deterministic equations. The agreement improves by increasing the number

of replicas and fine-tuning the anastomosis coefficient Γ as done in [41]. The value
given
R in Table 1, Γ = 0.145, produces the best agreement between the integer part
of p̃(t, x) dx calculated by solving the deterministic equations and by ensemble
averages of the stochastic process [41].

10. Conclusions. In our recent papers [4] and [41] we have explored the behavior
of a stochastic angiogenesis model, and of its possible deterministic approximation.
In this model, the tips undergo a stochastic process of tip branching, vessel extension
and anastomosis whereas TAF is described by a reaction-diffusion equation with a
sink term proportional to the tip flow.
In [4] the empirical measure describing the tip distribution had been assumed
to satisfy a “law of large numbers” for any single replica of the process, i. e. the
classical “propagation of chaos” assumption (see e.g. [40], and references therein),
so that it admits a position-velocity density which is shown to satisfy a nonlin-
ear integro-differential equation of a Fokker-Planck type, coupled with a reaction-
diffusion equation for the TAF concentration, in which the stochastic tip flow has
been replaced by its mean field approximation, deriving from the tip mean density.
On the other hand, in [41] we have solved numerically the stochastic model for
many realizations (independent replicas of the system). Numerically calculated
velocity fluctuations have revealed that they do not decay even as the number
of initial vessel tips increases. This shows that the stochastic model is not self-
averaging and therefore we cannot use the “propagation of chaos” assumption to
derive a mean field deterministic approximation of the stochastic model. The main
reason being that anastomosis eliminates many vessel tips, resulting in the fact that
there never are enough tips for a law of large numbers apply on a single replica. The
vessel network has shown to be quite different for different replicas of the stochastic
process.
However by re-examining the derivation given in [4], we conclude that the same
deterministic description holds for vessel tip densities calculated by averaging over
replicas. The deterministic description consists of a reaction-diffusion equation for
the TAF concentration coupled to a Fokker-Planck type equation for the vessel tip
density. The latter contains a birth term corresponding to tip branching and a
death integral term corresponding to anastomosis or tip merging. The coefficient
of the latter term has been fitted by comparison with the stochastic description:
optimal selection produces a good fit for the evolution of the total number of tips.
For the averaged deterministic reaction-diffusion system, boundary conditions
have been proposed in [41], which describe the flux of vessel tips injected from a
primary blood vessel in response to TAF emitted by the tumor and the tip density
eventually arriving at the tumor. Numerical solution of the model in a simple
geometry shows how tips are created at the primary blood vessel, propagate and
proliferate towards the tumor and may or not reach it after a certain time depending
on the parameter values. This is consistent with known biological facts.
Actually nontrivial additional investigation is required for a rigorous derivation
of the deterministic approximation of the relevant empirical measures and their evo-
lution equations. Anyhow we wish to convey a general message elicited by the pro-
posed angiogenesis model: in stochastic models containing birth and death processes
in addition to Brownian motion (Langevin equations), the death processes may pre-
clude reaching the large number of individuals required to have self-averaging and
64 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

a deterministic description based on the “propagation of chaos”. Nevertheless, de-

terministic equations for macroscopic densities and fluxes may follow from using
ensemble averages over a large number of replicas.
The significant consequence concerns the variance; though the mean behavior
can be described by the same PDE, the case of self-averaging does not carry any
variance; but the variance cannot be ignored in the case of ensemble averaging,
which implies the use of confidence bands in predicting the evolution of a real vessel
network! The authors are well aware of the limits of their own analysis, but they
wish to stimulate more attention to the mathematical issues raised by this important
biomedical problem. Only accurate models can support medical intervention for
prevention and cure. Simulations are surely useful as a first insight, but therapies
(optimal control) require accurate mathematical models, validated by comparison
with real data (inverse problems - statistics of random geometric structures).
Apart from the specific application we have been dealing with, in this paper
methodological contributions have been given for a sound mathematical modelling
of stochastic vessel networks: a) the use of stochastic distributions, and their mean
densities, describing the vessels - random objects of Hausdorff dimension 1; b) re-
duction of vessel distributions to integrals over time of tip distributions - random
objects of Hausdorff dimension 0; c) use of a Langevin model for vessel extensions,
thus leading to the required regularity for the existence of vessel mean densities [1].

Acknowledgments. The authors acknowledge the contribution of the anonymous

Referees for a significant improvement of the manuscript. This work has been sup-
ported by the Spanish Ministerio de Economı́a y Competitividad grant MTM2014-
56948-C2-2-P. VC has been supported by a Chair of Excellence UC3M-Santander
at the Universidad Carlos III de Madrid. It is a great pleasure to acknowledge
fruitful discussions with Daniela Morale of the Department of Mathematics of the
University of Milan.

REFERENCES

[1] L. Ambrosio, V. Capasso and E. Villa, On the approximation of mean densities of random
closed sets, Bernoulli, 15 (2009), 1222–1242.
[2] A. R. A. Anderson and M. A. J. Chaplain, Continuous and discrete mathematical models of
tumour-induced angiogenesis, Bull. Math. Biol., 60 (1998), 857–900.
[3] C. Birdwell, A. Brasier and L. Taylor, Two-dimensional peptide mapping of fibronectin from
bovine aortic endothelial cells and bovine plasma, Biochem. Biophys. Res. Commun., 97
(1980), 574–581.
[4] L. L. Bonilla, V. Capasso, M. Alvaro and M. Carretero, Hybrid modelling of tumor-induced
angiogenesis, Physical Review E, 90 (2014), 062716.
[5] P. Bremaud, Point Processes and Queues. Martingale Dynamics, Springer-Verlag, New-York,
1981.
[6] M. Burger, V. Capasso and D. Morale, On an aggregation model with long and short range
interactions, Nonlinear Anal. Real World Appl., 8 (2007), 939–958.
[7] M. Burger, V. Capasso and L. Pizzocchero, Mesoscale averaging of nucleation and growth
models, Multiscale Modeling and Simulation, 5 (2006), 564–592.
[8] V. Capasso, Randomness and Geometric Structures in Biology, in Pattern Formation in
Morphogenesis. Problems and Mathematical Issues (eds. V. Capasso, M. Gromov, A. Harel-
Bellan, N. Morozova and L.L. Pritchard), Springer, Heidelberg, 2013.
[9] V. Capasso and D. Bakstein, An Introduction to Continuous-time Stochastic Processes, 3rd
edition, Birkhäuser, Boston, 2015.
[10] V. Capasso and D. Morale, Stochastic modelling of tumour-induced angiogenesis, J. Math.
Biol., 58 (2009), 219–233.
 TUMOR-INDUCED ANGIOGENESIS 65

[11] V. Capasso, D. Morale and G. Facchetti, The Role of Stochasticity for a Model of Retinal
Angiogenesis, IMA J. Appl. Math., 77 (2012), 729–747.
[12] V. Capasso and E. Villa, On the geometric densities of random closed sets, Stoch. Anal. Appl.,
26 (2008), 784–808.
[13] P. F. Carmeliet, Angiogenesis in life, disease and medicine, Nature, 438 (2005), 932–936.
[14] P. Carmeliet and R. K. Jain, Molecular mechanisms and clinical applications of angiogenesis,
Nature, 473 (2011), 298–307.
[15] P. Carmeliet and M. Tessier-Lavigne, Common mechanisms of nerve and blood vessel wiring,
Nature, 436 (2005), 193–200.
[16] A. Carpio and G. Duro, Well posedness of a kinetic model for angiogenesis, Nonlinear Anal-
ysis; Real World Applications, 30 (2016), 184–212.
[17] N. Champagnat and S. Méléard, Invasion and adaptive evolution for individual-based spatially
structured populations, J. Math. Biol., textbf55 (2007), 147–188.
[18] M. Chaplain and A. Stuart, A model mechanism for the chemotactic response of endothelial
cells to tumour angiogenesis factor, IMA J. Math. Appl. Med. Biol., 10 (1993), 149–168.
[19] M. Corada, L. Zanetta, F. Orsenigo, F. Breviario, M. G. Lampugnani, S. Bernasconi, F.
Liao, D. J. Hicklin, P. Bohlen and E. Dejana, A monoclonal antibody to vascular endothelial-
cadherin inhibits tumor angiogenesis without side effects on endothelial permeability. Blood,
100 (2002), 905–911.
[20] S. L. Cotter, V. Klika, L. Kimpton, S. Collins and A. E. P. Heazell, A stochastic model
for early placental development, J. R. Soc. Interface, 11 (2014), 20140149, Available from:
http://rsif.royalsocietypublishing.org/content/11/97/20140149.
[21] R. F. Gariano and T. W. Gardner, Retinal angiogenesis in development and disease, Nature,
438 (2004), 960–966.
[22] J. Folkman, Tumour angiogenesis, Adv. Cancer Res., 19 (1974), 331–358.
[23] J. W. Gibbs, Elementary Principles of Statistical Mechanics, Yale Bicentennial Publications,
Scribner and Sons, New York, 1902.
[24] H. A. Harrington, M. Maier, L. Naidoo, N. Whitaker and P. G. Kevrekidis, A hybrid model
for tumor-induced angiogenesis in the cornea in the presence of inhibitors, Mathematical and
Computer Modelling, 46 (2007), 513–524.
[25] R. K. Jain and P. F. Carmeliet, Vessels of death or life, Sci. Am., 285 (2001), 38–45.
[26] S. Karlin and H. M. Taylor, A Second Course in Stochastic Processes, Academic Press, New
York, 1981.
[27] N. V. Mantzaris, S. Webb and H. G. Othmer, Mathematical modeling of tumor-induced
angiogenesis, J. Math. Biol., 49 (2004), 111–187.
[28] D. Morale, V.Capasso and K.Ölschlaeger, An interacting particle system modelling aggrega-
tion behavior: From individuals to populations, J. Math. Biol., 50 (2005), 49–66.
[29] K. Oelschläger, On the derivation of reaction-diffusion equations as limit dynamics of systems
of moderately interacting stochastic processes, Probab. Theor. Relat. Fields, 82 (1989), 565–
586.
[30] M. J. Plank and B. D. Sleeman, Lattice and non-lattice models of tumour angiogenesis, Bull.
Math. Biol., 66 (2004), 1785–1819.
[31] M. Hubbard, P. F. Jones and B. D. Sleeman, The foundations of a unified approach to
mathematical modelling of angiogenesis, Int. J. Adv. Eng. Sci. and Appl. Math., 1 (2009),
43–52.
[32] P. E. Protter, Stochastic Integration and Differential Equations, Second Edition, Springer-
Verlag, Heidelberg, 2004.
[33] G. G. Roussas, A Course in Mathematical Statistics, 2nd edition, Academic Press, San Diego,
CA, 1997.
[34] M. Scianna, L. Munaron and L. Preziosi, A multiscale hybrid approach for vasculogenesis and
related potential blocking therapies, Prog. Biophys. Mol. Biol., 106 (2011), 450–462.
[35] A. Stéphanou, S. R. McDougall, A. R. A. Anderson and M. A. J. Chaplain, Mathematical
modelling of the influence of blood rheological properties upon adaptative tumour-induced
angiogenesis, Math. Comput. Modelling, 44 (2006), 96–123.
[36] C. L. Stokes and D. A. Lauffenburger, Analysis of the roles of microvessel endothelial cell
random motility and chemotaxis in angiogenesis, J. Theor. Biol., 152 (1991), 377–403.
[37] S. Sun, M. F. Wheeler, M. Obeyesekere and C. W. Patrick Jr., A deterministic model of
growth factor-induced angiogenesis, Bull. Math. Biol., 67 (2005), 313–337.
66 BONILLA, CAPASSO, ALVARO, CARRETERO AND TERRAGNI

[38] S. Sun, M. F. Wheeler, M. Obeyesekere and C. W. Patrick Jr., A multiscale angiogenesis

modeling using mixed finite element methods, Multiscale Model. Simul., 4 (2005), 1137–1167.
[39] K. R. Swanson, R. C. Rockne, J. Claridge, M. A. Chaplain, E. C. Alvord Jr and A. R. A.
Anderson, Quantifying the role of angiogenesis in malignant progression of gliomas: In silico
modeling integrates imaging and histology, Cancer Res., 71 (2011), 7366–7375.
[40] A. S. Sznitman, Topics in propagation of chaos, in École d’Été de Probabilités de Saint-Flour
XIX–1989 , Lecture Notes in Math., Springer, Berlin, 1464 (1991), 165–251.
[41] F. Terragni, M. Carretero, V. Capasso and L. L. Bonilla, Stochastic model of tumor-induced
angiogenesis: Ensemble averages and deterministic equations, Physical Review E , 93 (2016),
022413.
[42] S. Tong and F. Yuan, Numerical simulations of angiogenesis in the cornea, Microvascular
Research, 61 (2001), 14–27.

Received November 23, 2015; Accepted April 13, 2016.

E-mail address: luis.bonilla@uc3m.es
E-mail address: vincenzo.capasso@unimi.it
E-mail address: mariano.alvaro@uc3m.es
E-mail address: manuel.carretero@uc3m.es
E-mail address: filippo.terragni@uc3m.es